Remote Sensing Using AVHRR

Editor: Kevin Gallo Global Cl imate Labora to ry National Climatic Data C e n t e r Federal Building Asheville. N.C 2880 1 U.S.A.

Dynamic Stratification of the Landscape of Mexico: Analvsis J

of Vegetation Patterns Observed with Multitemporal Remotely Sensed Images

Franz Mora* Department of Forestry, University of Illinois. Urbana, IL 61820, U.S.A.

Louis R. Iverson USDA Forest Service, 359 Main Road, Delaware, OH 43015, U.S.A.

Abstract

Rapid deforestation i n Mexico, when cozrpled w i t h poor access to current and consistent ecological information across the col lntry underscores the need for a n ecological classification sys t em that can be readily updated and n e w data beconze available. In this s tudy , regiorzal vegetation resources i n Mexico were evaluated us ing remotely sensed information. Mtllt i temporal Global Vegetation l n d e x { G V I ) data from Advanced V e r y High Resolution Radiometer images provided ecological infornzation at regional scales b y being interpreted as plzenological patterns ofvegetat ion product iv i ty and seasonality. Principal component analysis o n G V I lnonthly composites identified spatial and temporal vegetation patterns, reducing their variation to f ive phenologically m e a n i n d u l components . S i x t y land- cover and natural vegetation classes zuer: then derived via unsupervised classification f rom the f ive principal components . Addi t ional phenoloipical i~zfontzation (e.g., onset and peak of greenness, periods of growth) was obtained for each class. These data , along wltlz seasonality measures (e.g., szrnzmer us . w in ter peak ofgreenness) were used as criteriafor group ing similar vegetation and land-cover types in to a classification for Mexico.

Introduction

The natural landscape of Mexico has changed drastically in the last 30 years. Deforestation has reached unprecedented levels, a situation that is not likely to change significantly during thls decade. According to the Mexican National Forest Inventory, 6 million

hectares of all types of forest land were converted to other uses between 1964 to 1984, representing a loss of 25% from the total forested areas reported in the first national inventory (Inventario Nacional Forestal 1964, 1991). The estimated annual deforestation rate was 0.71%) between 1980 and 1990, and projections suggest a deforestation rate of 0.55% for the current decade

* Current address: Center for Advanced Land Management Information 'Technologies, University of Nebraska at Lincoln, 113 Nebraska Hall, P.0.Box 880517, Lincoln, NE 68588-0517. U.S.A

73 Ccncnrto Iiitcrnntic~~ini, Vo1.12, No.2, June 1997 Published by Geocarto l~~ternatioilal Centre, G.P.O. Box 4122, Hong Kong.

(Inventario Nacional Forestal 1991). Additionally, non official estimates indicate that the trend in Mexico is one of the highest among tropical ecosystems, with a deforestation rate approaching 10% m tropical forested regions (Myers 1993). Due to the high rate of landscape conversion, better decision making is needed for improved land-use policy development and implementation.

Presently, efforts to slow the rate of deforestation by implementing new land-use regulations are limited by a lack of information, especially at the regional level. A recently erected environmental law (Ley Forestal 1990) mandates that land managers apply ecological principles to forested areas, watersheds and national parks, via management and conservation plans. The long and short-term effects of disturbance are being evaluated primarily on the basis of regional data.

Regional ecological information for Mexico is not always available to decision makers. The most recent cartographic products related with natural resources in Mexico were obtained through visual interpretation of Landsat images from the 2970s. Mexican cartographic resources are, in general, outdated, inaccurate, and sometimes nonexistent (Hough 1993), and digital sources of illformation are just now becoming available (INEG1 1991). Because of the lack of map products, remotely sensed satellite data seem to be the most current source of information. Only such data can support regional landscape analysis at this time (Townshend et nl. 199:l). Remotely sensed data also provide statistically valid estimates of ecological parameters over large areal extents (Botkin et al. 1984).

Multitemporal analysis of ecological attributes at the regional scale is essential in understanding the dynamics of natural landscapes. Analysis of spatial and temporal variations of landscape features, such as vegetation and climate, is necessary to understand modifications in landscape structure resulting from land-use change, deforestation, and perturbation effects in general. Since the effects of lai~dscape modification are revealed at several temporal and spatial scales over ecological gradients, it also is necessary to develop a framework analysis that integrates temporal variations and spatial heterogeneity of ecological landscape features. A landscape ecological classification defines such a framework, and is enabled when recurrent temporal and spatial patterns are stratified into ecologically meaningful landscape units.

There are several sources of satellite data that allow an objective analysis of ecological variables at different scales, from ecosystems to landscapes (Wickland 1991). Among them, the Advanced Very High Resolution Radiometer (AVHRR) from the National Oceanic Atmospheric Administration (NOAA) has two desirable characteristics: it has a high temporal frequency and its products are available at several scales of observation. Because of the daily frequency of AVHRR observations,

there is a high probability of obtaining at least one cloud-free image in every part of the world each week or two using a composting technique (Oaring et al. 1989). Image composting allows the comparison of images collected over a sequential period (e.g., 7,14, or 30 days) where the maximum measurement (e.g., normalized difference vegetation index, NDVI) is retained to represent the conditions observed during that particular period (Holben 1986).

AVHRR images are available at three grid cell sizes: (1) Global Vegetation Index (GVI), with a cell size of 16 krn at the equator (GVI is a NDVI composite); (2) Global Area Coverage (GAC), with 1.1 by 4.4 km at nadir; and (3) Local Area Coverage (LAC), with 1.1 by 1.1 km cell size at nadir (Kidwell 1990). Different cell sizes in AVHRR products allow analysis of temporal trends of ecological variables associated with these observations at different spatial scales (Malingreau and Belward 1992), though each of these scales might be too coarre for local analysis.

The primary characteristics of AVHRR and its relationships with ecological variables have been discussed elsewhere (Box et a!. 1989; Cihlar et al. 1991; Goward and Dye 1987; Goward 1989; Goward et al. 1991; Hastings and Emery 1992; Holben 1986; Roller and Colwell1986). The most important AVHRR-derived variable for ecological applications is the NDVI, which has been shown to be well suited for vegetation analysis.

Previous work has shown that multitemporal NDVI images are useful for analyzing spatial vegetation patterns from regional to continental scales (Goward et al. 1985,1987; Justice et al. 1985; Townshend et al. 1987; Tateishi and Kajiwara 1991; Tucker et al. 1985), and for assessing vegetation dynamics (Nelson et al. 1987, Nelson 1986). In addition, when a stratification according to some ecological criterion is needed, vegetation dynamics can be described using AVHRR (Eidenshink and Hass 1992). Practically, the imaging frequency and compositing process makes it possible to describe regional vegetation o n a seasonal (phenological) basis (Lloyd 1991).

Environmental applications of AVHRR include land- cover mapping vegetation dynamic studies, tropical forest monitoring, fire risk assessment, vegetation production and biophysical parameter estimation (Ehrlich ct al. 1994). However, multitemporal analysis of vegetation activity using remotely sensed data has become one of the main applications of AVHRR images. Using NDVI-AVHRR images, land-cover classes can be separated in a multitemporal space according to phenological, seasonal, and latitudinal variations in vegetation (Ehrlich et al. 1994).

Principal component Analysis (PCA) and Time Series Analysis are used frequently to capture the seasonal variation in multitemporal datasets (Townshend et al. 1985; McGwire et al. 1992, Eastman and Fulk 1993, Reed el ul. 1994). PCA can be used to reduce the

dimensionality of the multitemporal dataset, i.e., reducing the number of variables ("dimensions") in the analysis. However, the real potential of PCA lies in its ability to identify the true number of linearly independent vectors in the original matrix (Davies 1986). These linear vectors are interpreted as series of new and uncorrelated "components," which are combinations of the original variables (monthly NDVI values).

Principal components are usually computed from eigenvectors of the covariance matrix between variables, This results in orthogonal representations of variation (using and orthogonal rotation method) since the covariance matrix is symmetric. Although other methods of rotation can be used (Richman 1986), orthogonal rotations capture the periodicity inherent in the data (Goodman 1979). The components define linear combinations of original variables where their respective eigenvectors are proportional to the fraction of the variance of the original dataset accounted for by eachcomponent. Usually, the first component accounts for the majority of the variability; subsequent components explain residual (but still significant) variance, capturing all of the details in their modes of variation.

Because the resulting components reflect a combination of monthly NDVI values, their meaning is more complex than the original variables. Although the resulting interpretation of principal components derived from multitemporal NDVI analysis remains a matter of judgment, these can be related to seasonal vegetation activity, and their modes of variation can be mapped. Further, the spatial representations (images) of each component represent a series of latent images or trends that would be nearly impossible to detect by direct examination of the data (Eastman and Fluke 1993).

In this context, PCA assumes a meaningful interpretation of the components obtained. Previous studies are consistent in giving (at least for the first component) interpretation to PCA results. Generally, the first component is a measure of the non-seasonal and locational variability of vegetation, while the other components obtained can be quantitatively related to the "green up" or "brown down" of vegetation (Ehrlich et al. 1994). A classification into land-cover classes, using the components obtained with PCA, integrate the phenological variations of vegetation, while classifications using the original bands d o not necessarily do so. However, the amount of seasonal variability captured by PCA depends on the number of months represented in the multitemporal dataset, year of observation, variability in the vegetation activity, and noisiness of the scenes (e.g., subpixel cloud contamination, sensor anomalies, etc.). In short, PCA is highly scene dependent, and results should be analyzed in the proper context.

The objective of this paper is to show how the application of remotely sensed data (specifically GVI images) can be used for multitemporal landscape analysis in Mexico using PCA. The application of PCA in multitemporal analysis of vegetation activity results in valuable phenological information that can be represented in several ways, including land-cover classifications. Although a previous land-cover classification for Mexico was developed using AVHRR composites (December and May) from NOAA-11 (Evans et al. 1992), this study constitutes one of the first attempts to capture the seasonal component in a vegetation/land-cover classification for Mexico. A recent effort using GVI data and PCA for Mexico showed that the seasonal component of vegetation can be captured (Turcotte et al. 1993) though the modes of variation were not analyzed in the pl~enological context.

The identification of recurrent multitemporal and spatial patterns in GVI images (and other AVHRR products) should result in improved sources of vegetation information for an ecological classification. The application of this technique to this specific dataset does not attempt to obtain a definitive classification for the country, but rather to illustrate how phenological information can be interpreted in an ecological sense. Better multitemporal vegetation information (than GVI) will be available through AVHRR (i.e., the NOAA/ NASA Pathfinder AVHRR Land dataset) and other platforms (i.e., EOS) in the future, but this and other contributions using PCA techniques should help "standardize' the methods of approach.

Methods

Sources of Information A subset of GVI (NDVI) monthly averaged

observations (January to December) was extracted for Mexico from the global dataset developed by the Construction Engineering Research Laboratory (CERL) Environmental Laboratory in Champaign, IL (CERL, undated). The CERL-Global dataset contained monthly composited data for 45 months from April 1985 to May 1989; these were averaged into 12 monthly values. Although these monthly composited data may, at times for some portions of the globe, be contaminated by continuous cloud cover or spurious sensor artifacts, our examinatios of the 4-year averaged data did not show this to be problematic for Mexico. The NDVI is calculated from channels 1 and 2 of daily GAC data; its values are scaled (minimum value of 0 and maximum of 65) to represent the data, from no vegetation productivity to maximum vegetation productivity, in 8-bits (Kidwell, 1990). The cell size in all GVI images contained in the CERL-Global dataset were arbitrarily resampled to 4 minutes 48 seconds (0.08 decimal degrees) per grid cell, though the source data were 8.64

minutes per grid cell ) (Tateishi and Kajiwara 1991).

Data Analysis The original CVT data were mapped by month to

obtain a phenological characterization associated with the Mexican vegetation during 1985-89. With the analysis of these data, the following phenological metrics about vegetation could be mapped: (1) maximum photosynthetic level, (or maximum monthly GVI score) for each pixel; and (2) minimum photosynthetic level (or minimum monthly GVI score) for each pixel (see Lloyd 1992).

PCA was applied to the GVI dataset to determine the statistical dimensionality of seasonal variations in the landscape. PCA also can reduce the total variation in the original 12 (monthly) GVI bands on an annual basis and produce components that are highly related with vegetation productivity and seasonality. The scree test, which estimates variation accounted for each component, was used as an aid in choosing the number of components to retain in subsequent analyses. An orthogonal rotated solution was applied to calculate the respective principal component scores. Respective component values were calculated for each resulting principal component and used as a set of new ecological variables for an unsupervised classification. The principal component procedure was implemented using the PRINCE algorithm in ERDAS software (ERDAS 1990).

For pattern identification, an unsupervised classificatioi~ approach was preferred because there was no preconceived number or types of classes that define the landscape units. Sixty preliminary "greenness" classes were derived from the five principal component values using an iterative self-organizing data-analysis technique (ISODATA), which is a spatial classificatory algorithm (ERDAS 1990). The ISODATA algorithm was selected because it gives better results than other methods (e.g., statistical clustering using parallelepiped or minimum-distance methods) while identifying clusters inherent in the data. The 60 classes represent a stratification of the spatial and spectral variation captured by the principal GVI components across Mexico.

Original GVI values for each class were plotted by month to visualize the temporal pattern obtained with PCA. Each class had a characteristic "phenological signature" that represented photosynthetic activity during a correspondent period of growth. These signatures can be used to obtain a phenological classification of vegetation activity, according to Lloyd (1991). Even when phenology in vegetation is associated in an agricultural context (planting, fruiting, and harvest), it also has been defined as the "study of the timing of recurring biological events, the causes of their timing (due to biotic and abiotic forces), and interrelationships among species." Seasonality also can

be defined in terms of the "occurrence of certain obvious biotic and abiotic events or groups of events within a definite limited period of the astronomic year" (Lieth 1974).

There are several phenological (timing of onset and peak) and seasonal (summer-winter difference in vegetation activity) variables that can be observed directly using multitemporal NDVI signatures for each class. The onset of greenness is observed at the month when there is a significant departure of previous GVI values (generally indicating an acceleration of the photosynthetic activity); the peak of greenness is determined when the maximum GVI value is reached for that class. Senescence in vegetation also can be also observed when greenness declines, and the end of the growing season can be identified when declining GVI values reach levels similar to those observed at onset. The duration of tile growing season (in months) was identified by comparing the occurrence of onset and senescence dates in the signatures.

Three categorical maps (onset, peak, and duration of growing season) were obtained by reclassifying the unsupervised classes according to this phenological information, A more complete phenological characterization was then obtained by masking the onset, peak, and duration maps with the original GVI values. Vegetation index values for the "onset month" give the photosynthetic activity level at the beginning of the growing season, while GVI values at "peak month" give the photosynthetic level at the peak of the growing season.

Each of these variables was combined into a raster data base that describes the vegetation phenology variation in Mexico. A correlation analysis among all phenological factors was performed on the raster data to explore redundancy in the phenological set. Pearson's correlation coefficients were obtained using ARC/ INFO'S correlation command (ESRI 1995). Correlation coefficients were squared to obtain an estimate of the proportion of variance that can be explained by each phenological factor as a function of each other.

The 60 unsupervised classes were interpreted and labeled using vegetation types and land-cover categories reported in the "Land Use and Vegetation" map prepared by the Instituto Nacional de Estadistica Geografia e Informiitica (INEGI 1980). This map at the scale of 1 : l million, was derived from visual interpretation of photographic Landsat products dating from the late 1970s. Obviously, there are several difficulties with this approach in labeling unsupervised classes. One problem is caused by the difference in dates between the GVI data and the creation of the map. Another problem surfaced because the patches corresponding to particular vegetation types in the INEGI map did not always match with the distribution of any particular class or set of classes in the GVI- derived data. In this case, the classes were named

according to the most similar vegetation and land- cover category.

images. These maps show predictable patterns in the

vegetation, according to GVI variations: ( 1 ) the desert ecosystems in Baja California and northeastern Mexico never show much photosynthetic activity; (2) the tropical regions in the Yucatan Peninsula are highly photosynthetic for much of the year; (3) the conifer forests running along the "Sierra Madre Occidental" (western side of the county), which are largely pine forests, have relatively high GVI scores throughout the year but especially in the summer months; (4) the agricultural regions of central Mexico reach peak greenness during the summer months; (5) the deciduous

Results and Discussion

Spatial and temporal trends in GVl The 12 monthly-averaged GVI images captured the

annual and spatial variation of vegetated features in the landscape of Mexico during the period 1985-89 (Fig. 1 ). In addition, a summarization of original monthly GVI data shows maximum and minimum photosynthetic activity (Figs. 2a & 2b), which aids in interpreting the patterns elucidated by the satellite

February

r-kl

I April -J lune

August September

November December I

October " ... .. -- - -. . . ...... ----'I a

CVI level m >&I432 Multitemporal GVI-NDVI Variation i n Mexico

>&%a3 Albers Conic Equal-Area >a344 - >o?cas

/ I . ............. - ._ . . . .- -.. .......... - - 1

Figure 1 Multitemporal variation in vegetation activity. Monthly global vegetation index (GVI) from January to December.

vegetation in central Mexico also shows a phenological pattern similar to agriculture, and reaches maximum photosynthetic values during summer months; and (6) maximum photosynthetic activity is similar between tropical and temperate forests, though minimum photosynthetic activity is substantially higher ill the tropical regions (Figs. 1,2a & 2b).

Principal Components Analysis A scree test in the PCA Indicated that five principal

components could be retained while explaining 94.7% of the variation in GVI. Thus PCA thereby reduced the

dimensionality of 12 monthly GVI images into five phenologically meaningful components.

Scores for the five principal components were plotted by month to show the structure of each principal component (Fig. 3). This graph shows the PC structure in terms of the original variables (monthly GVI's) that are more associated with each principal component. PC1 is relatively constant across the year, whereas PC2 to PC5 have considerable seasonal variation. PC2 peaks in August and has its minimum in March, whereas PC3 peaks in December and has its minimum in May. PC4 and PC5 are bimodel in nature (Fig. 3).

(c) Uegetation Productivity -

w

(el Photosynthetic Level a: Onset

(b) Minimum Photosynthetic ~evel

(f) Photosyntheti ' el at Peak

0

0 onai >oaos

--..- -

Figure 2 Phenological variation derived from 12 monthly GVI scores (scaled from 0 tol). (a) Maxi~nuln and (b) lninilnuln photosynthetic activity, (c) vegetation productivity, (d) vegetation seasonality, (e) onset and (f) peak of photosy~~thetic activity.

The PCA also provided, wit11 PC1 and PC2, the capability to map indirect measures of vegetation productivity and direct measures of vegetation seasonality (Figs. 2c & 2d). PC1 values showed an increasing trend ill a north-south direction, highly associated with the neotropical-tropical pattern of vegetation distribution, i.e., very low values in the northern part of the country and the Penninsula of Baja California (where deserts and semideserts occur), and considerably higher values in the Yucatan Peninsula (where tropical vegetation occurs). As such, PC1 is highly related to vegetation productivity, as has been shown previously (Townshend ot 01. 1985; Coward and Dye 1987).

The pattern observed for PC2 showed that the country can be divided into two general areas according to vegetation seasonality. The first, or summer season growth zone (scores of > 0.3 in Fig. 2d), occurs in the lowlands of central and northern Mexico, as well as in the northern portion of the Yucatan Peninsula. In these areas, the associated vegetation is highly deciduous (deciduous selvas and agriculture). The second, or winter season growth zone (score < 0.3 in Fig. 2d), occurs in the extreme south, parts of the "Sierras," and in deserts of the northern portion of the country where perennial vegetation (deserts, conifers, and rain forests) dominates the landscape.

Classification of vegetation types A classification based on five principal components

rather than on the original 12 months of GVI data offers several advantages. First, the seasonal variation in vegetation identified through PCA can be included in the classification scheme. Vegetation types and land-

cover classes can be described in productivity and seasonal components in addition to phenology. Additionally, subsequent multivariate analysis, which uses the components as a set of new ecological variables, can be performed for an ecological classification without temporal autocorrelation effects, and can be assumed to be statistically independent, i.e., multicollinearity effects on successive monthly observations are removed. Since each component obtained with PCA is not only orthogonal but also statistically independent, the classification will not be temporally autocorrelated and the differences among classes will be maximized with respect to the new phenological information.

The unsupervised classification on the principal components resulted in a definition of 60 land units (classes) associated with different seasonal vegetation patterns. In addition, since ISODATA is an algorithm that takes into account the spatial context, i.e., classifying neighboring pixels, the definition of classes primarily relies on the maximization of the variance between classes that are in close proximity, and minimizes the variation within the classes. Thus, the land units identified using the components classified with this algorithm are temporally and spatially homogeneous.

To rank the vegetation types as they are separated by each of the first three principal components, the PCA coefficients scores and class means of each of the 60 classes identified with ISODATA were plotted in the feature space plot created by PCA (Fig. 4). PC1 apparently corresponds well with the annual accumulated NDVI (Townshend et a/ . 1985), which differentiates the greenest features of the landscape from less green features (for Mexico, it clearly

1

0.75

0.5

0.25

0

-0.25

-0.5

-0.75

1

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Figure 3 Principal compollent structure according to the variation in monthly factor scores coefficients.

differentiates selvas from deserts, Figs. 2a & 4). PC1 has been previously associated with annual integrated NDVI values in North America, which, in turn, are associated indirectly with functional ecological variables such as gross and net primary productivity, and actual evapotranspiration rates (Goward and Dye 1987).

PC2 is the primary component which captures the temporal variation of natural vegetation (Figs. 2d & 4). High PC2 values represent highly seasonal vegetation types that reach peaks of greenness durlng from July to September (e.g., low deciduous selva and rainfed

agriculture). Lower values are mostly associated with seasonal vegetation that reaches peak of greenness from November to April i.e., coniferous forests. Thus, the maximum amplitude in PC2 variation is accounted for by summer-winter differences in vegetation growth (Fig. 4a).

PC3 to PC5 are seasonal components that differentiate features that are not associated with the "natural" seasonal variation identified by PC2. PC3 identifies areas that are associated primarily with agricultural zones irrigated during the winter season (Fig. 4b). PC4 and 5, though not shown graphically,

(a) High Productivity

Low Productivity

Winter Growth S u m m e r Growth 4 PC2 b

0 ~~arcocnulcsccnt Dcscrt

0 Sarcophyllous Dcscrt

Chaparral

0 Fl;q;itc ( Pinc oak

Forcst

Forest

Secondary vcgctation Subtropical Forcst

1 ~;i;caulcscenl Grassland Oak Forest Subpcrcnn~al Sclvos 0 ~~~C;phyllous ( -"~maulipan Mountainous Pcrcnn~al

1 hornscrub I .rsophy,lous Sclvas

0 ~gr icul ture ) Submountainous Deciduous Irrigated matorral Selvas

(b) High Productivity

Low Productivitv

Figure 4 Feature space plot evaluated for class means in (a) PC1 and PC2 and (b) PC I and PC3.

identify agricultural areas that are irrigated in May. ecosystems at the biome level due to the scale of GVI In this analysis, landscape units represent primarily observations. These 60 spectral classes also can be

Table 1 Land-cover and vegetation types reported on INEGI map associated with the unsurpervised classification of Mexico.

2 Sarcopl~yllor~s desert Desert vegctalion, dominated by leaf sarcopliyllo~~s s ~ ~ c c ~ ~ l e n t plants (Ago1.e) and Desert, rnatorral rosetolilo 9.629 5 pcrenriials slirubs ( l ~ h r i r r r i r i .vp, Yrrcco salhla). costero, matorrat desertico

roselolilo 3 Crassica~~lescet~t Desert vegetation, dominated by large cactus (stem-succ~~lent) ((.'anrcgiu grparrfea. Desert, ~nalorral crasicaule, 12,472 7

desert Feroorf r r r ~vislizer~ii) cylindropuntia and platyopuntia, and perennial shn~bs cardonal ((:ercidiusr microph~~llunr).

1 Micropllyllous Desert vegetation, dominated by ephemeral l~erbaceo~~s species (growing winter Semideserts, n~atorral desertico 10.390 6 desert season) and a low density o f perennial plants. Larren tridrritoto and 17ranscria microtilo, ~natorral espinoso

[hrrrio.sa.

CLASS

5 Agriculture Rainfed agriculture. 51.498 28

1 Sarcocaulescent Desert vegetation, dominated by large sarcocaulescent lrees (Iltrrsera micraphylb Desert, cactus scn~b, ~natorral 6,761 4 desert and Jafmpl~o cmerea) with s ~ ~ c c u l e ~ ~ t plants (Idrio colrrnrriarir), evergreen shn~bs xerofilo

(Iarrea frickrilafa), and decidooos shrubs (Jarmphn cirneota).

DESCRIPTION LAND COVER

6 Chaparral Plant c o m ~ n ~ ~ n i t y o f shrubs and scn~bs developed in a Mediterranean climate with 1,628 I species resilant to lire (C)rrcrcar rp. A~lcnosfonra sp ) and annual grasses (Arcfosfaphylas rp and (:crcircarprr.r sp)

7 Mesquite s l ln~b Se~i~idesert vegetation dominated by lrees with small leaves, generally mesquite Woodland, savanna, mesquite- 4,555 2 (Prf~.vopis q). but Acacia ,sp and dercidirrnr sp are also conlmon. grassland

8 Grassland Annual grasses. Pastizal, zacatonal.sabana 10.9 15 6

OTHER TERMINOLOGY

9 Tamanlipan Semidesert shn~bs and low forest formations, wit11 thorny and deciduous species 'n~orn forest, t l~orn woodland, 9.608 5 thornscnlb (Acacia sp, C'crcidiurrr sp, I.cacol~l~!~llrrnr rp. I'rri.sopi.v .rp). lnatorral espinoso tamaulipeco

10 Sub~l~o~~ntainous Semidesert shn~bs and deciduous woodlands (Helietta pan'folia, Neoprirrglec~ Yl~orn woodla~id, matorral, 3,033 2 n~atorral irrfcgrili,lia, Cordia hai.r.rieri, Acacia sp ma1 Lerrcaphyllrrnr sp). Considered a matorral sobmontano, selva

transition zone between semidesert matorrals, deciduous selvasa. and oak forest. baja espinosa.

Area

[I$ ha]

I I Pine oak forests Conifer telnperate forest dominated by species of pine (Pkruv s11) and oaks Boreal foresl, bosq~~e de 10.737 6 (Qvcrcirv .rp). Characteristic species vary with the geographic region. coaiferas, pinar encinar

% Mexico area

12 Pine forests Conifer temperate forest dominated by pines. Conifer forest, pinar, bosque de 7.713 4 pino

13 Oak forests Temperate forest do~n i~~a ted by oaks. Encinar, bosque semihlimedo 2,073 I de montana

14 Monntaino~~s Temperate deciduous forest on n~esic I~illsides don~inated by deciduous broad- Cloud forest,~nontane rain 3,158 2 mesopl~yllous leaved species (Digclhardlin nrcxicorra. Jrrg1mr.v olachmro, Liqrridanrhar forest, lemperate deciduoos forests styracflrma, and Osfria ~'irgir~iarra) and wit11 evergreen (Pinrrr sp) species forest, mo~oitain mesie forest,

dominating the lower strata. bosque mesotilo de montana

15 Deciduo~~s dry selva Selva ill warm dry clin~ates with l ~ i g l ~ l y deciduous (6-8 monlhs o f growing season) I)ec~duous seasonal foresl. 12.267 7 and semideciduous species (Brrr.xra b~si losra sp, Jacarafia rrrexicarro, tropical deciduous forest, selva Iponroco sp. Pscerlohnnrhnr sp, Cbcropia sp. Ccndrela sp, elc) and with a dense baja caducifolia. bosque thicket-like understory. tropical decidrro

16 Second growtli Vegetation t l~at results f ron~ secondary succesio~~ in areas which have been cleared vegetation for agricultural or grazing purposes.

17 Subtropical forests Subtropical thorny shrubs, matorrals, and forests that share characteristics o f each. Semi-evergreen seasonal 5.062 3 Considered a transition zone between temperate foikst and matorrals. forest,tropical d e c i d ~ ~ o ~ ~ s forest, Characteristic species are Iponronea sp. Burrera sp, Ey.seirhardfia p i ~ ~ . ~ t a c h i a . bosque tropical subcaducifolio and Acacia sp.

18 Perennial selvas Three-storied selva, in tropical clin~ates, dominated almost exclusively by Rain forest, tropical evergreen 9.216 5 perennial species with straight, unbranched, buttressed lm~ iks rising 50-60 m from forest,selva alta perennifolia, the foresl floor (Aspido.vpenria nregolocorpon, Bro.vinrrm alica,vmm~. Llialrm bosque tropical perennifolio, priurrerrsc. Ternrirralin arrrazariia, Swicfertia niacrophylla, Vmhysia guafen~alcnsi.~, selva alta sie~npre verdc, selva Alchonrca lafifi~lia, Alihertia cd,rIi.r, Bclolia canihellii Brrnrclia persinri1i.r. Bvrsem ombrolila siempre verde anranrha). V ie second stratunl contains a continuous canopy o f branched trees o f 25-10 m. The third stratum contains small lrees 10-20 m tall.

19 Subperennial selvas Selva with a dominance o f subperennial trees and a well-defined dry season, and Montane rain forest, se~ii i- 8.780 5 with less overdue precipitation. Superficially . i t has similar composition as evergreen seasonal foresl. perennial selvas, but has only two and ocassionally one layer o f trees. "Dominant" bosqi~e tropical sr~bcad~tcifolio, species are Actmniirnr graveokn.v, Benrorrllia J7amnica. Brir.~inrrini dica.~fnmi, bosqi~e deciduo semihumedo Orirrrclia persinrilis. Ceiha perifo~tdra.

1 20 Irrigated agriculture Irrigated agricultural lands. 2.962 2 - -

" ~ e l v n is a term in Spanish which distinguishes between the highly diverse communities of trees developing in the lowlands, in contrast to bosque that describes the communities of trees in the highlands and mountains.

specifically related to 20 vegetation and land-cover classes defined apriori via the INEGI (1980) map. Thus, statistically valid areal estimates (Table 1) and a map depicting the distribution of the resulting vegetation and land-cover classes can be obtained (Fig. 5).

The results of vegetation and land-cover mapping (Fig. 5) shows that patterns of vegetation distribution are highly consistent with those observed in multitemporal GVT monthly variation and PCA. Substantial differentiation can be obtained by classifying contrasting phenological vegetation types (e.g., conifers, rain forests and deserts), but vegetation and land-cover classes are less evident when they occur within s~milar phenological groups (i.e., semidesert type). PC3 also allows the identification of highly contrasting phenological zones such as irrigation areas that probably would be less evident without multitemporal information.

Results of the classification also show that most of

the Mexican landscape can be classified as "natural" (69%) by summing the classes dominated by natural vegetation. The remaining portion is classified as cultural (agriculture, irrigation, and secondary growth), most of which (28% of the total area of Mexico) is predominantly agriculture. The highest proportion of natural landscape categor~es is in deserts, deciduous selvas, grasslands, and conifers, each of which occupies only about 6 and 7% of the total area in Mexico (Table 1).

The characteristic GVI variation of each class during the year defines a multitemporal spectral signature for each ecosystem or land-cover unit identified in this analysis (Fig. 6). Ecological information that is not evident without temporal analysis can be extracted from these signatures, and when combined with PCA which d i s t i n ~ u i s h e s the classes, allows the characterization of the landscape in several ways not previously possible (Figs. 2e, 2f & 4). Phenological attributes such as peak and onset of greenness, levels

90°W'00" 105*00'00" IOQ'OO'W. 95'W'OO'

10'00'00

9'W'OO'

L B'M'W"

Land Cover and Vegetation types

u rarcocaulescenl desen I grassland dec~duous dry selva

iarcoohyllous desert I larnalilipan thornscrub 7 second growtt vegetation

' crassicaulerenc desert I s~bmounlainous matorral rubtrop~cal forests

m microphyllous desert olne oak foresls perennial selvar

2 agriculture pine forests subperennial selvas

I I/ chaparral I I oak forests irrigated agriculture

mesqulte shrub ( mountainous merophyll forests

Figure 5 Land cover and vegetation types.

of minimum and maximum photosynthetic activity, and beginning and duration of the growing season along with seasonality measures, can be obtained for specific locations (Fig. 2, Table 2). Subsequently, all of these attributes can be used as multivariate variables to classify landscape units in ecological terms.

However, not all of the phenological information derived with this analysis is completely independent. Correlation analysis among phenological variables showed that phenological attributes can be redundant in capturing their respective spatial patterns of distribution (Table 3).

Squared correlation coefficients among phenology

Deserts

variables showed that considerable variance (67%) can be explained between vegetation productivity (PC1) and maximum and minimum photosynthetic levels, as well as with the photosynthetic level at peak. This reinforces the idea of PC1 as a vegetation productivity component. More than 80% of the variance in both photosynthetic levels at peak and onset can be explained if the maximum GVI registered is used rather than combining PCA and the unsupervised classification. However, maximum GVI values during the year give no indication of the months in which these occur.

Photosynthetic levels of vegetation activity, particularly peak and onset of the growing season, are

, , , , . . . . . , , , JAN FEE MAR APR MAY JUN JUL AUG SEP OCT NOV DEC

Irrigated Agriculture

0

0.00

JAN FEE MAR APR MAY JUN JUL AUG SEP OCT NOV DEC

Semideserts

O-" 6

JAN FEU MAR APR MAY JUN JUL AUG SEP OCT NOV OEC

Deciduous Selvas and Grasslands

0.00 m 1 JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC

Conifer Forests Selvas (Tropical Rainforest) 0.60 I 0.60 I

JAN FEE MAR APR MAY JUN JUL AUG SEP OCT NOV DEC JAN FEE MAR APR MAY JUN JUL AUG SEP OCT NOV DEC

Figure 6 Spectral signatures for individual classes identified with ISODATA. Individual spectral classes (lines wtthin each group) are associated with land covers and vegetation types and grouped 011 a phcnological basis.

Table 2 Phenological classification of vegetation activity using multitemporal GVI and principal component analysis.

Phenological Characterization of vegetation in Mexico* based on Multitemporal GVI imagery and PCA

Phytophenological attribute: NDVI definition:

I) Minimum level of photosynthetic activity Minimum GVI value 2) Maximum level of photosynthetic activity Maximum GVI value 3) Beginning of growing season Month at which "Onset" was

recorded 4) Peak of growing season Month at which "Peak" was recorded 5) Photosynthetic level at the beginning Onset (GVI) value

of the growing season 6) Photosynthetic level at the peak of the Peak (GVI) value

growing season 7) Annual vegetation productivity index PC I value 8) Seasonality index for natural vegetation PC2 value 9) Seasonal indexes for nonnatural vegetation PC3, PC4 and PC5 values

(Agriculture) *~Mod?fied from Lloyd (1 99i

Table 3 Pearson's correlation coefficients (squared) for phenological variables derived from rnultitemporal GVI imagery and principal component analysis in Mixico.

I 2 3 4 5 6

Vegetation Productivity 1 .OO

Vegetation Seasonality 0.240 1.00

Maximum photosynthetic level 0.672 0.058 1.00

Minimum photosynthetic level 0.672 0.017 0.436 1.00

Photosynthetic level at onset 0.518 0.032 0.828 0.348 1 .OO

Photosynthetic level at peak 0.672 0.068 0.884 0.462 0.792 1 .OO

highly correlated with each other and with vegetation productivity, but seem to be highly independent from seasonality measures (Table 3). PC2 is completely independent information about vegetation phenology. None of the other phenological variables are correlated with vegetation seasonality, indicating that PCA gives real (not apparent) variation in seasonal vegetation (Fig. 7). ,The correlations among the phenological variables can be identified visually by their respective spatial patterns (Fig. 2).

The probable causal factors of temporal and spatial variations of individual classes, and consequently of landscape units, can be established using additional information such as climate and topographic variables. If such associations were found, the spatial patterns of landscape units could be predicted by a GIs database, with the results of the multitemporal analysis used as a valuable source of temporal and spatial vegetation

patterns for Mexico. Such an analysis and GIs implementation is currently in progress.

Conclusions

In this study, principal component analysis ~f multitemporal GVI images of Mexico summarized the vegetation variation in the landscape. Current improvements in the acquisition and resolution of multitemporal vegetation data, via remote sensing, will continue to help alleviate the problem of lack of ecological data for the country. With the application of PCA onmultitemporal (LAC, GAC or GVI) data, several phenologically meaningful components can be obtained that represent the total annual variation of vegetation dynamics. Since PCA captures nearly all of the temporal variation in the multitemporal GVI dataset, it seems well suited for the seasonal vegetation analysis of the

landscape of Mexico. PCA also offers several advantages in subsequent multivariate analysis (i.e., identifying classes) since the new variables obtained are statistically independent. Variance reduction techniques also can significantly reduce the CPU time in calculations as well as disk space.

The new components derived from this analysis can be interpreted as phenological variables and used to characterize the landscape in terms of productivity, seasonal variations in natural vegetation, and seasonal variations in "non-natural" vegetation types (e.g., agricultural lands). Additionally, the interpretation of multitemporal signatures for every class allows the

characterization of every landscape unit in phenological attributes. The use of spatial classifiers in the definition of spectral classes also allows the capture of spatial heterogeneity of every landscape unit.

Since PCA (as with other variance-reduction techniques) is highly scene dependent, this interpretation of components and classes is valid only for the 1~mdscapeinMexico as described here. However, similar interpretations can be obtained for other regions.

Acknowledgments

The authors gratefully acknowledge Elizabeth Cook,

Onset of Greenness

Peak of Greenness

January

ebruary

Aarch

L .,

I U. r m luly

August

September

October

November

December

Two seasons

1 3 months

4 months

5 months

~

1 j

i

r '

Duration of Greenness _ _ . _ __. _ _ ._ _. .A_I..-___-..-...-. -. -. - -- - Figure 7 Seasonal variation of vegetation activity in Mexico. (a) onset, and (b) peak of growing

season, (c) duration of growing season identified with GVI.

6 months

7 months

8 months

9 months

10 months

11 months

2 seasons

Susan Egen-McIntosh and Zuliang Zhu for improving the quality of the manuscript at early stages. Special thanks to the Illinois Natural History Survey for providing the equipment and facilities to conduct this research. This research was conducted while the first author held a Fulbright-CONACYT-IIE scholarship for his M.S. degree at the Department of Forestry, University of Illinois, Urbana-Champaign. This research was partially funded by the Tinker foundation of the University of Illinois at Urbana Champaign.

References

Botkin, D.B., Estes J.E., MacDo~iald R.M., and Wilson M.V. Studying the earth's vegetation from space. Biosciencc 34: 508-514. 1984.

Box, E.O., Holben B.N., and Kalb V. Accuracy of the AVHRR vegetat ion index a s predictor of biomass, pr imary productivity and net CO, flux. Vegetatio 80: 71-89. 1989.

Cihlar, J., St-Laurent L., and Dyer J.A. Relation between the normalized difference vegetation index and ecological variables. Retnote sensing of Environinent 35: 279-298. 1991.

Construction Engineering Research Laboratory. Glohal GRASS 1. CD of 50 global coverage maps. Cook College Remote Sensing Center, Rutgers University, New Brunswick, NJ.

Davies, J.C. Statistics and data onalysis in gcolo,yy. 2nd edition. John Wiley and S(IIIS, New York. 1986. 646 pp.

Eastman, J.R. and Fulk, M. Long sequence time series evaluation using standardized principal components. Photogramn~etric Enginccrin~ and Ren~ote Sensing 59: 1307- 1312. 1993.

Eideiisliink, J.C. and Hass, R.H. Analyzing vegetation dynamics of land systems with satellite data. Geocarto Intcri~ationul 7: 53-61. 1992.

Ehrlich, D., Estes, J.E., and Singh, A. Applications of NOAA- AVHRR l k m data for environmental monitoring. lntrrnational / o u ~ ? ~ N ! ifRr7notc Srnsirfg 15: 145- 161. 1994.

ESRI. Environmental Systems Research Institute, Inc. ARC/ INFO user's manual. Environmental Systems Research Institute, Inc. Redlands, CA. 1995.

ERDAS. Earth Resources Data Analysis lnc. Field guide. Version 7.5. ERDAS. Atlanta, Georgia. 1990. 410 pp.

Evans, D.L., Zhu Z., Eggen-Mclntosh S., Garcia-Mayoral P., and Ornelas d e Anda J.L. Mapping Mexico's forest lands with advanced very high resolution radiometer. USDA Forest Service Research Note 50-367: 1 -4. 1992.

Goodman, D. Applications of eigenvector analysis in the resolution of spectral pattern in spatial analysis and temporal ecological sequences. In: G.P. Patil and M. liosenzweig (eds.) Contcniporury Quantitativr Ecolo<y,y rwd Related Econo~ncfrics. pp. 139-155. International Co-operative, Fairland MD. 1979.

Coward, S.N., Tucker C. I., and Dye D. North American vegetation patterns observed with NOAA-7 advanced very high resolution radiometer. Vixrtutio 64: 3-14. 1985.

Coward, S.N. and Dye, D.C. Evaluating North American net primary prod~~ct iv i ty with satellite observations. At/vuncrs ill Space Research 7: 165-174. 1987.

of North American and South American biomes from AVHRR observations. Geocarto International 2: 27-40. 1987.

B Coward, S.N., Satellite bioclimatology. \oiirnal of Climate 2: 710-

720. 1989. I Goward, S.N., Markham, B., Dye, D.C., Dulaney, W., and Yann, I .,

J Normalized difference vegetation index measurements from the advanced very high resolution radiometer. Remote Sensing of the Environment 35: 257-277. 1991.

Hastings, D.A and Emery, W.J. 1992. The advanced very high resolution radiometer (AVHRR): a brief reference guide. Photograinmetric Engirzeerln,? and Remote Senslng 58: "1 183- 1188.

Holben, B.N. Characteristics of maximum-value composite images from temporal AVHRR data. International Journal of Remote Sensing 7: 1417-1434. 1986.

Hough, H. Mapping Mexico's Future. Earth Observation Magazine (Feb): 50-52. 1993.

Instituto Nacional de Estadistica Geografia e Informatica. Carta dc uso del suelo y vegetacibn escala 1: 1,000,000. Instituto Nacional d e Estadis t ica Geografia e Informatica, Aguascalientes, Mexico. 1980.

Instituto Nacional d e Estadistica Geografia e Informatica. Catalogo d e Productos INEGl. Instituto Nacional d e Estadistica Geografia e Informatica, Aguascalientes, MPxico. 171 p. 199'1.

Inventario Nacional Forestal. Primer Inventario Nacional Forestal, Direction General del INF; Subsecretaria Forestal. Secretaria de Agricultura, Mexico. 1964.

Inventario Nacional Forestal. lnventario Nacional Forestal de Gran Vision, Direccinn General del INF. Subsecretaria Forestal, Secretaria de Agricultura, Mexico, 60 p. 1991.

Justice, C.O., Townsliend, J.R.G., Holben, B.N., and Tucker, C.J. Analysis of tlie phenology of global vegetation index usii~g meteorological satellite data. lnternational Journal of Remote Sensrng 6: 1271-1318. 1985.

Kidwell, K.B. Gloi~ul z~egetation lniiex guide. US Department of Commerce, NOAA Administration, Satellite Data Service Division, Washington DC. 1990.

Ley Forestal. Primera Ley Forestal, Diario de MPxico. Mexico. 1990.

L~eth, H. Purposes of a phenology book. In: Lieth, H . (ed.) Plienology aiiil scusorrality niorieling, pp. 3-19. Springer-Verlag, New York. 1974.

Lloyd, D. A phenological classification of terrestrial vegetation cover using shortwave vegetation index imagery.Internationa1 journal of Remote Sensing 11: 2269-2279. 1991.

Malingreau, J.P., and Belward, A.S. Scale considerations in vegetation monitoring using AVHRR data. International jo~iri~al of Reniot~ Srnsing 13: 2289-2307. 1992.

McGwire, K.C., Fairbanks, D.H., and Estes, J.E. Examining regional vegetation associations using multi-temporal AVHRR imagery. Proceedings of American Society for Photogram~iletric Engrnrerlng and Reniotr Striising, American Congress of S~rrveyrng and Mapping, March 1992, Washington DC. p p 304-313. American Society for Photogrammetry and Remote Sensing. Bethesda, Maryland. 1992.

Myers, N. Tropical forests: the main deforestation fronts. Coward, S.N., Kerber, A,, Dye, D.G., and Kalb, V. Compar i so~~

Environmental conservation 20: 9-16. 1993.

Nelson, R. Ident ifying deforestat ion in Brazil using multiresolution satellite data. International lournal of Remote Sensing 7: 429-448. 1986.

Nelson, R., Horni~ig, N., and Stone, T.A. Determining the rateof forest convers io~~ in Matto Crosso, Brazil, using Landsast MSS and AVHRR data. Internutional lournal of Remote Sensing 8: 1767-1784. 1987.

Ohring, G., Callo, K., Gruber, A,, Planet, W., Stowe, L., and Tarpley, J.D. Climate and global change: characteristics of NOAA satellite data. EOS 70: 889-901. 1989.

Reed, B. C., Brown, J.F., VanderZee, D., Loveland, T.R., Merchant, J.W., and Ohlen, D.O. Measuring phenological variability from satellite imagery. lo~trnal of Vegetation Science 5: 703- 714. 1994.

Richman, M.B. Rotation of principal component analysis.journa1 cf Clirriatology 6: 293-335. 1986.

Roller, N.E.G., and ~ o l w e l l , J.E. Coarse resolution satellite data for ecological surveys. Biosciencc 36: 468-475. 1986.

Tateishi, R., and Kajiwara, K. Land cover monitoring in Asia by NOAA GV1 data. Geocartr~ International 6: 53-64. 1991.

Townshend, J.R.G., Goff, T.E., and Tucker, C. Multitemporal dimensionality of images of normalized difference vegetation index at continental scales. IEEE Transactions on Geoscience and Remote Sensing GE-23: 888-895. 1985.

Townshend, J.R.C., Justice, C.O., and Kalb, V.T. Characterization and classification of South American land cover types using satellite data. International journal o j Remote Sensing 8: 1189- 1207. 1987.

Townshend J., Justice, C.O., Lei, C. W., Gurney, C., and McManus, J. Global land cover classificatiol~ by remote sensing: Present capabilities and future possibilities. Remote Sens~ng of the Environment 35: 243-255. 1991.

Tucker, C.J., Townshend J.R.G., and Coff T.E. African land cover classification using satellite data. Science 227: 233-250. 1985.

Turcotte, K.M., Lulla, K., and Venugopal, G. Mapping small- scale vegetation changes in Mexico. Geocarto International 4: 73 -8 5 . 1993.

Wickland, D.E. Mission to planet earth: The ecological perpective. Ecology 72: 1923-1933. 1991.